Thread: FPN math

Also now that I've looked at it a bit more you're version relies on integers bigger than a float, as you may have guessed I will be attempting long doubles later so I'd like to use a version that does not rely on larger integers because that opens up the problem of what to do when there is no such integers, while I can program array type integers I still struggle with the division part which would be essential at the top of your function, I would also need to program a log10 etc variant for those array integers, all that still excludes the possibility of not enough memory anyway.

Originally Posted by awsdert

Also now that I've looked at it a bit more you're version relies on integers bigger than a float, as you may have guessed I will be attempting long doubles later so I'd like to use a version that does not rely on larger integers because that opens up the problem of what to do when there is no such integers, while I can program array type integers I still struggle with the division part which would be essential at the top of your function, I would also need to program a log10 etc variant for those array integers, all that still excludes the possibility of not enough memory anyway.

Take a look at strtod() implementation from glibc... it uses multiple precision (MP) arithmetic. So you need integers with precision greater than the targeted floating point type. You only need integers to hold 2 times the precision of integer type (which is always bigger then the precision of the floating point type). This could be worse for convertions from char to "float", for example, where the expoent is between -126 and 126 (for normalized floats) - perhaps you'll need a MP with 256 bits (to be safe), to do those shifts. But it's not a big deal in terms of memory usage.

As you've seen, log10 and pow10 is not a big deal to implement...

With my test, if you are using GCC or CLANG, you can use __int128 type for long doubles.